Uncertainty-guided active transfer for robotic suction grasping via evidential learning

Robust uncertainty quantification is crucial for accurate scene understanding and reliable decision-making in robotic suction grasping. Although existing data-driven grasp detection methods demonstrate effectiveness, they often fail to quantify predictive uncertainty and suffer from low sample efficiency during simulation-to-reality transfer. To address these limitations, we propose an uncertainty-guided active transfer framework that leverages evidential deep learning to jointly estimate aleatoric and epistemic uncertainties for evaluating model confidence and environmental ambiguity. Building on this quantification, an asynchronous training pipeline is introduced, driven by a hybrid sampling strategy that iteratively selects both highly uncertain and highly certain target-domain samples. This strategy optimally balances domain exploration with model exploitation, facilitating discriminative training to bridge the simulation-to-reality gap while enhancing sample efficiency. Furthermore, we design an evidential feature fusion mechanism that synergistically integrates the predictive distributions of suction seal quality and the object’s center of mass with their associated uncertainties. By incorporating these uncertainty estimates into the final grasp extraction pipeline, our framework ensures informed decision-making and robust execution. Comprehensive experiments demonstrate that the proposed uncertainty-driven approach significantly enhances learning efficiency, achieving consistently reliable grasping performance across diverse and complex real-world scenarios.